Make code match comments - use 64bit float

Author: mankoff

xarray/coding/variables.py

@@ -237,7 +237,7 @@ def _choose_float_dtype(dtype, has_offset):
         # Sensitivity analysis can be tricky, so we just use a float64
         # if there's any offset at all - better unoptimised than wrong!
         if not has_offset:
-            return np.float32
+            return np.float64
     # For all other types and circumstances, we just use float64.
     # (safe because eg. complex numbers are not supported in NetCDF)
     return np.float64